search

msracver / Deformable-ConvNets

Deformable Convolutional Networks
MIT License
4.05k stars 957 forks source link

Could you explain why mask generated by Sigmoid in mdcn is doubled? #246

Open JrPeng opened 5 years ago

JrPeng commented 5 years ago

In your paper, the mask of each layer is generated by a sigmoid with its value initialized as 0.5. However, the mask is doubled with 1.0 as initial value in this implementation. Could you explain the mismatch between the code and the paper? Or, could you release the results of 1x mask and 2x mask? Thank you.

zheningy commented 5 years ago

My guess is that the normal mask for every sample point should be 1. However, the sigmoid function is 0.5 in zero. Thus it is required to be doubled to become 1.